GPU Runners

Your production GPUs should run your product, not your test suite.

NVIDIA RTX PRO 6000, RTX 5090 and RTX 4090, in the GitHub Actions workflow you already have. Same platform, same agents opening pull requests that make your runs faster.

the lineup

Three GPU options with more coming soon

Every GPU runs Ubuntu 24.04 with the same agents watching your workflows. Click a label to copy it, then paste it after runs-on:. Prices are per minute.

StarSling GPU runners, with the label and per-minute price of each GPU
LabelGPUHostPer minute
NVIDIA RTX PRO 6000 defaultThe default GPU. Large-model and memory-heavy CI workloads.4 vCPU · 16 GB · 100 GB$0.05922 / min
NVIDIA RTX 5090Fast inference tests and model evals.4 vCPU · 16 GB · 100 GB$0.03022 / min
NVIDIA RTX 4090Everyday GPU tests at the lowest cost.4 vCPU · 16 GB · 100 GB$0.02522 / min
Coming soonNVIDIA H100Maximum throughput for large-model CI workloads.Coming soonComing soon

Tell us which GPU you want and what you plan to run on it, and we will open access for your org.

For the latest labels and pricing, see the StarSling docs.

Weighing a fleet of your own? Our guide to self-hosted GitHub Actions runners has the instance cost, the operating surface and the fork-code exposure that comes with running them.

running gpu jobs

Spend GPU minutes on the jobs that need them

  • Trigger the heavy suite deliberately

    Put the long CUDA builds and model eval suites behind a pull request label or a workflow_dispatch, and let every push run the CPU suite. The jobs that need a GPU are usually a small share of the pipeline, and they are the ones worth choosing.

  • Run the full matrix nightly

    A schedule: trigger at a quiet hour covers the combinations a pull request does not, and the results are waiting when the team starts. Billing runs from job start to finish, so a nightly matrix costs the minutes it runs.

  • Filter by path

    Scope the GPU workflow with paths: so it fires on the source it actually covers. Our guide on path filters has the allowlist versus denylist reasoning and the traps.

what you get

The hardware, the billing, and the agents

  • NVIDIA RTX PRO 6000 to 4090

    The same platform, backed by an NVIDIA GPU. Pick it in the runs-on: label, and the job runs on it in less wall time than the box it runs on today. The GPU is dedicated to that job for as long as it runs, so nothing reclaims it partway through your suite.

  • You pay for run time

    Billing runs from job start to finish, rounded up to the nearest minute, and queue time is not billed. Concurrency is unlimited, and your first month includes 2,000 free minutes.

  • Agents that keep tuning

    StarSling reads your real runs and opens pull requests that shard suites, fix cache configuration and cut critical path. The runners get faster after the swap, not just at it.

the rest of the family

Same YAML, different silicon

Give production its GPUs back.

Tell us your org and the GPU you want, and we will open GPU access. CPU runners are available today if you want to move the rest of the pipeline first.